The Accord Created an Auditor's Chair.
It Didn't Create the Measuring Stick.

On September 29, 2026 the President and six companies building frontier models signed the White House Accord on Super Intelligence. The document asks each company to build four layers of controls and audits. Three of the four are internal: internal controls, an internal team, and a committee of the company's own board. The fourth asks each signatory to partner with an independent external auditor or evaluator to assess whether those controls are operating as intended. The accord never defines who qualifies as that evaluator, what standard the assessment measures against, whether results must be published, or what happens if they are bad. Independence has two dimensions: who performs the assessment, and who defines what is being measured. The accord addresses the first and leaves the second unanswered.

On September 29, the President and six companies building frontier models signed the White House Accord on Super Intelligence. It runs about three hundred words.

Four layers of controls. Only the third leaves the building. It asks each signatory to "partner with an independent external auditor or evaluator."

That phrase is the entire external check in the document. The accord never says who qualifies to be one.

The day after the signing, the Federal Trade Commission confirmed a probe that reaches the frontier labs and the evaluator a signatory had named three weeks earlier.

Read the accord closely. What it does not say is more specific than what it does.

Three of the four layers of control are the company checking itself

The four layers, in the accord's own order:

1. Internal controls monitoring the capabilities and alignment of its models during training and deployment.
2. An internal team to ensure those controls, monitoring, and detection are operating and that issues are remediated.
3. An independent external auditor or evaluator to carry out independent assessments of whether they are operating as intended.
4. An "independent committee of the board of directors" to oversee and receive the reports. Independent directors. Notably, not independent of the company.

Layers one, two, and four are the company checking itself. Layer three is the only one that leaves the building. Each signatory commits to "partner with an independent external auditor or evaluator to carry out independent assessments of whether the controls, monitoring, and detection are operating as intended."

Read that last phrase again. Operating as intended.

Intended by whom.

The standard the external auditor measures against is the developer's own statement of what the system was supposed to do. The company writes the specification. The company builds to it. The auditor then assesses whether the company met the company's own target.

That can be independent assurance. What it is not is independent measurement against an independently established standard. The auditor may be external. The measuring stick is still internal.

Independence has two dimensions: who performs the assessment, and who defines what is being measured. The accord addresses the first and leaves the second unanswered.

Four words doing all the work, none of them defined

"Independent external auditor or evaluator" is the entire external check in the document. The accord never says who qualifies.

What the accord does not specify about that evaluator:

• Accreditation. Who qualifies to perform the assessment.
• Conflicts. What relationships disqualify a candidate.
• Appointment and removal. Who selects the evaluator, and on what terms it can be replaced.
• Payment. Who pays, and whether that creates a dependency.
• Access. What the evaluator is guaranteed to see.
• Publication. Whether the result reaches anyone outside the company.
• Standard. What the assessment measures against.
• Consequence. What follows if the assessment is bad.

There is no regulator, no penalty, no revocation, and no floor beneath which a company cannot fall. The document creates the auditor chair and specifies neither who may sit in it nor what they measure with.

The signatories acknowledge this. The accord closes by noting that "over time, it may make sense to codify these steps into laws or regulations." That is the industry pointing at the gap in its own document.

The mechanism is still under discussion

Dario Amodei, who runs Anthropic and seventeen days earlier called for pacing the frontier, granted on the day of signing that "the technology has very real risks." Then he added: "And, you know, the mechanism, how we address those risks is still under discussion."

The mechanism is still under discussion. That is an accurate description of the document he had just signed. The accord records the commitment before the mechanism has been specified.

Seventeen days earlier he had published something more specific. Writing about interpretability, the science of understanding what happens inside a model, he wrote: "Despite all the progress, we still only understand a tiny fraction of what goes on inside these models."

Hold those two statements together. The people building these systems say they understand a tiny fraction of what happens inside them. The accord asks an external evaluator to assess whether those systems are operating as intended.

Against what baseline.

An evaluator who can be selected can be un-selected

The same September essay, published seventeen days before the signing, proposed the arrangement the accord gestures at. Amodei called for "embedded evaluators," writing that each frontier company should commit to "giving ongoing, employee-like access to a team of embedded third-party evaluators," naming one such organization directly, and committing his own company unilaterally.

Employee-like access is his phrase, and it is the honest one. It describes the actual arrangement: an evaluator selected by the lab, embedded in the lab, operating with the access the lab grants, assessing the lab against the lab's own stated intent.

That may well be the most practical arrangement available today. Employee-like access is not itself evidence of independence, and neither is physical separation. Independence requires rules: who selects the evaluator, who pays it, what conflicts disqualify it, what access it is guaranteed, whether findings can be suppressed, and under what conditions it can be removed. The accord specifies none of them. Without rules governing appointment, removal, access, payment, conflicts, and publication, the document provides no basis for determining how independent an evaluator remains once selected. An evaluator who can be selected, embedded, and granted access can also be un-selected.

The accord does not address this, because the accord does not define the auditor at all.

The evaluator is in the information request

On September 30, the Federal Trade Commission confirmed it is investigating Anthropic, OpenAI, and other AI companies over the risk their technology poses to consumers. The agency is examining whether their conduct runs afoul of the FTC Act, the statute that governs unfair or deceptive practices. Reporting indicates it is preparing civil investigative demands, the instrument used to compel documents and executive testimony.

One detail in the confirmation deserves more attention than it received. An FTC spokesperson told CBS News the agency "plans to request information from the companies, including the nonprofit research group METR."

METR evaluates frontier models for risk. Its name is Model Evaluation and Threat Research, and it is pronounced meter. It is the organization a signatory named when he proposed who should check his own company's work. Seventeen days before the signing, in the same essay where he proposed embedded evaluators, Dario Amodei named METR directly as the kind of third party that would verify adherence to safety commitments.

The significance is not that an information request proves METR is insufficiently independent. It does not. An evaluator may receive a request precisely because it holds relevant independent evidence. The significance is that the evaluator now sits inside the evidentiary perimeter of the same federal inquiry examining the companies it may evaluate. That makes the accord's failure to define evaluator independence more consequential, not less.

The agency was not a late arrival. OpenAI's own chief global affairs officer put it in the room. "We were at the White House on Tuesday," he said. "The FTC was actually there when we announced those safety commitments."

The FTC spokesperson said the probe was first opened over the summer. What is not established is when METR entered its scope. The request became public on September 30, one day after the signing.

FTC Chair Andrew Ferguson had made the structural argument nine days before the signing. Speaking on Fox News on September 20, he argued that the two companies should not be allowed to alarm Washington into passing rules they are already positioned to satisfy. "That is how companies build a moat around their businesses to make sure that people can't compete against them," he said.

He also said the government is not without tools today. "It's not like the government has no tools to protect Americans," he said. "We have the product liability laws. We have the consumer protection laws."

That is the sitting chair of the agency now examining them, arguing that safety regulation can become a competitive moat when incumbent companies help shape the rules. He is not disputing that the risks are real. He is making a claim about who benefits when the people who build the technology also define the standard it is measured against.

The accord does not answer the FTC chair, because the accord never defines the auditor. It asks each company to partner with an independent external evaluator and leaves every qualifying term blank. If the evaluator can be named by the lab, embedded in the lab, and asked for its files in the same inquiry, the word independent is carrying weight the document never gave it.

This is worth stating plainly, because it is the part most of the coverage inverts. The role is not hypothetical. The accord names it, a signatory has named the organization he would put in it, and that organization is already doing the work. The problem is not that nobody is doing the job. The problem is that nothing specifies what the job measures against, and nothing requires the result to be published.

The word he never said

The day after the FTC investigation was confirmed, OpenAI's chief global affairs officer, Chris Lehane, went on television to answer for it.

Asked directly whether OpenAI wants to self-police, he did not hedge. OpenAI does not want to self-police, he said. "To be really clear, if there's one message, we do want to see national mandatory safety standards," he said. The interviewer acknowledged it, and he continued: "We want that to be passed in the law. We want legislation, and we are advocating for it."

That is a signatory of a voluntary accord saying, on the record, that voluntary is not what the company wants. Take him at his word. It is more than the document itself asks for.

Then listen for what the standard is made of. He reached for the strongest available precedent, unprompted: "the US created the airline industry. We put in place the FAA. The FAA became the global standard for air travel and safe air travel."

The FAA is the right shape. It is a public body that sets requirements manufacturers do not write for themselves, and on paper it certifies against them whether or not a manufacturer agrees.

He never said who would do that for AI. In the transcript as aired, the words third party, independent, evaluator, auditor, and verification do not appear at all.

Earlier in the same interview he described his company's product this way: "We have the most capable model in the world out right now. It also happens to be the safest model in the world."

The safest in the world, measured by the company that built it, in the same conversation where he calls for mandatory standards. Both statements can be sincere. That is the point. The absence of an external instrument is not a position the participants appear to be defending. It is the unresolved gap between the commitments they are making and the measurement regime those commitments require. Whose measuring stick is he using?

Asked how a company earns public trust, he said "you have to earn it. You have to show it, not tell it," and offered as proof that 1.2 billion people use the tools. Usage can establish adoption. It cannot establish that users have independently validated the system's safety or reliability.

He also mentioned, in passing, what his company shipped that week. "We first of all announced a new tool called dots," he said. "Think of it as an assistant in your pocket. It can save you a ton of time on your daily life, helping you with emails and managing daily life challenges."

OpenAI introduced dots on September 29, the same day the accord was signed. It is an agent that works across your mail, your calendar, and the apps you connect to it, and acts on your behalf.

To be fair to OpenAI, dots ships with real controls, and the announcement describes them plainly. Background research runs on tools "restricted to be read-only, which means that they can't send messages, change app content, or control your browser or computer." Your machine "and its contents stay separate unless you choose to connect it." Custom rules let you require approval or block actions outright. Sensitive tasks, the company says, "always stay with you." The company built all of this without being told to.

Then, directly from the same announcement page: "Dots can still make mistakes, so always review consequential work."

That is an honest disclosure, and it is also the distinction this whole document turns on. The controls are OpenAI's. The monitoring that can pause a dot is OpenAI's. The auto-review deciding what proceeds is OpenAI's. Internal controls, documented in detail. What no part of it provides is independently verified performance: an outside party establishing whether those controls do what the company says they do. The accord asks for exactly that and does not say who performs it.

And even a genuinely independent evaluator still needs to know what it is measuring.

A system doing the right thing on false information

Every hazard the accord enumerates is a system doing the wrong thing. Layer one asks companies to monitor their models "around areas like cybersecurity, biosecurity, and chemical threats, and to ensure that its models do not hack or access technical systems in unintended ways."

Hacking. Biosecurity. Chemical threats. Unintended access.

Every named hazard is a system doing the wrong thing. Cyber-physical harm is not among them, and neither is the failure below.

A system doing the right thing on false information is not in the document. Not excluded, not argued against, simply never named.

That omission matters more than the ones people are arguing about this week. A model that is confidently wrong and cannot tell does not require misalignment, bad actors, or a capability jump. It is the ordinary failure of a competent system, it is observable today, and nothing in the document aims at it.

Consider what that failure looks like when it reaches someone. You file a legitimate claim and it is denied. The record behind the denial is false and looks entirely genuine. Every rule was followed. Every permission was honored. Nobody lied.

You appeal. The same system reviews the appeal against the same record and confirms the original decision. Every check the process contains has now run twice and agreed with itself. The audit trail is clean. The outcome is still wrong, and nothing in the process that produced it was built to ask whether the underlying record was true.

Name the distinction that makes. Controls can verify that a system followed its authorized process. They cannot, by themselves, establish that the information entering that process was true. Procedural correctness and epistemic correctness, whether the process was followed and whether what went into it was true, are different measurements. The document explicitly asks for the first. It never establishes a corresponding measure for the second.

A larger concession than the coverage noticed

The accord is voluntary. It creates no regulator and imposes no penalty. Flanked by the executives who signed it, the President called it "almost like a constitution, in a way." Asked whether the deal was binding, he said: "I think it's morally binding."

A constitution is binding. That is most of what makes it a constitution. A binding rule has three things this one does not: someone empowered to check, a consequence for failing, and a way to lose the privilege. Speed limits have all three. This has none.

None of that is the interesting part. Voluntary commitments have moved industries before when the measurement underneath them was real.

The interesting part is that a document signed by the President and six frontier labs now says, in writing, that independent external assessment is necessary. That is new, and it is a larger concession than the coverage suggested.

Having said that, the document declines to say what such an assessment would consist of, who is qualified to perform one, or what scale it would be measured against. Those are answerable questions. I have set out what an independent measurement would have to satisfy in the Foundational Framework, and a reader is free to disagree with the conditions it proposes; what the field does not yet have is an agreed way to argue about them.

External evaluation without an external standard is not yet a measurement regime.

The evaluator chair has been created. It is occupied. The measuring stick has not been built.

Update, October 6: the seat that got staffed

Five days after this essay published, who selects and who pays were still blank. The White House staffed a different seat. The President named Jay Clayton, the sitting Director of National Intelligence, to chair a new Super Intelligence Force. The FTC Chairman and the Under Secretary of War for Research and Engineering serve as vice chairs, alongside the Director of the Office of Personnel Management, reporting to the President and Chief of Staff. The charter gives them 120 days, and directs them to "develop plans for responding to SI-enabled threats to our society, while preventing overregulation and regulatory capture that would stifle innovation and competition." To the charge that the companies are policing themselves, Clayton's answer was "we have a Federal Trade Commission, we have a Department of Justice." He added that he knows the leaders of both organizations and that "they're committed to protecting the American people in this regard." Enforcement after harm is not evaluation before deployment.

Frequently asked

What does the White House Accord on Super Intelligence require?

Four layers of controls and audits from every company training frontier models: internal controls monitoring capabilities and alignment, an internal team confirming those controls work, an independent external auditor or evaluator assessing whether they operate as intended, and an independent committee of the company's board receiving the reports. The accord is voluntary. It creates no regulator, imposes no penalty, and provides no way to lose standing under it.

Does the accord create an independent AI auditor?

It creates the role and leaves every qualifying term blank. The document asks each signatory to partner with an "independent external auditor or evaluator" without defining who qualifies, what accreditation is required, what conflicts disqualify, whether findings must be published, or what standard the assessment measures against. The chair exists. What the occupant measures with does not.

What is the difference between independent assurance and independent measurement?

Independent assurance means an external party assesses whether controls operate as the company intended. Independent measurement means scoring against a standard the company did not write. The accord provides for the first and not the second. Independence has two dimensions: who performs the assessment, and who defines what is being measured. An auditor can be genuinely external while the measuring stick stays internal.

What does "operating as intended" mean in the accord?

It means the external evaluator assesses performance against the developer's own statement of what the system was supposed to do. The company writes the specification, builds to it, and the auditor confirms whether the company met the company's own target. That is a second opinion on a self-selected question, not measurement against an independently established standard.

What is the difference between procedural correctness and epistemic correctness?

Procedural correctness asks whether a system followed its authorized process. Epistemic correctness asks whether the information entering that process was true. Controls can verify the first. They cannot, by themselves, establish the second. A system can follow every rule, honor every permission, produce a clean audit trail, and still reach a wrong outcome because the underlying record was false. The accord asks for procedural correctness only.

Why does evaluator independence need to be defined rather than assumed?

Because independence is structural, not a claim. It requires rules governing who appoints the evaluator, who pays it, what conflicts disqualify it, what access it is guaranteed, whether findings can be suppressed, and under what conditions it can be removed. Without those rules, a document can ask for an independent evaluator and provide no basis for determining how independent that evaluator remains once selected.

---

Disclosure: I develop independent measurement methods for AI agent governance and hold patent filings in that field. I have a commercial interest in the gap this essay describes. The accord's text, the FTC confirmation, and every quotation here are primary-source documents a reader can check without taking my word for any of it.